Discovering Online Shopping Preference Structures in Large and Frequently Changing Store Assortments

Abstract

The authors develop an attribute-based mixed-membership model of consumers’ preference for stockkeeping units in store assortments. The model represents the underlying “topics of interest” that drive shopping behaviors as probability distributions over product attributes. It overcomes several limitations of latent Dirichlet allocation topic models and is particularly useful for making preference predictions in large and frequently changing store assortments. The authors apply the proposed model to investigate topics driving browsing and purchase activities in an online deal marketplace of fashion products and explore how preference structures evolve over time. They find commonalities and differences in the topics that drive the browsing and purchase stages of online shopping processes. In general, browsing covers a broader range of product attributes than purchases. Consumers tend to browse products of premium positioning and/or deep discounts in the deal marketplace, but when purchasing, they tend to gravitate toward lower-tiered products at their original prices and modest depths of discounts. The authors illustrate how insights from the proposed model can be utilized to profile consumers based on their price preferences and to improve personalized product recommendations. They show that the model's performance is particularly strong in predicting preferences for new products that are not in the existing assortment.

Keywords

SKU-level models topic models fashion retailing assortment management personalized product recommendations machine learning

Many retailers carry large and frequently changing assortments. With the development of e-commerce, the assortment size of a typical online store or marketplace is even larger and growing steadily. Unlike for consumer-packaged goods, consumers hardly ever buy the same item multiple times in many other types of merchandise, such as clothes, footwear, bags, jewelry, fashion accessories, video games, music, movies, books, and electronics. Therefore, retailers selling such merchandise need to frequently update their assortments. In addition to large assortment sizes, the frequently changing nature of their assortments makes it difficult to predict consumers’ preferences and to make effective merchandising decisions.

The general approach proposed in the marketing literature to studying consumers’ purchase decisions in fairly large assortments is to translate individual stockkeeping units (SKUs) into relevant product attributes, and then model purchase decisions as driven by preferences for those product attributes (e.g., Fader and Hardie 1996; Ho and Chong 2003; Inman, Park, and Sinha 2008; Sinha, Sahgal, and Mathur 2013). As well articulated in the seminal work by Fader and Hardie (1996), an important advantage of modeling consumer preferences for product attributes is its ability to make preference predictions for new products that have yet to be included in the assortment. Nonetheless, prior SKU-level models were mostly constructed at the product category level. They were not meant to incorporate cross-category dependencies of purchase behaviors in the entire store assortment.

Recent developments in topic models for natural language processing have offered new tools for handling the scale of retail assortments. An analogy can be drawn between words and documents studied by natural language processing to purchased products (or SKUs) and a consumer's purchase history record in retailing data. Based on this analogy, Jacobs, Donkers, and Fok (2016, 2021) have adopted the latent Dirichlet allocation (LDA) topic modeling approach to studying purchase behaviors concerning the entire store assortments and investigated the underlying topics (which they called “motivations”) that drove purchase decisions. Their approach provides rich insights on consumers’ preference structures in terms of plausible motivating factors. For example, in the context of online grocery products, Jacobs, Donkers, and Fok (2016) identified underlying “motivations” such as a preference for ecofriendly products or low-fat products.

Nonetheless, the LDA models have several limitations which hinder their ability to assist managerial decisions facing many retailers, especially those with frequently changing assortments. The most serious limitation is that they cannot provide preference predictions for products that do not exist in the calibration data of existing assortments. In this study, we propose an attribute-based mixed-membership (ABMM) model, which is better suited for frequently changing store assortments. The proposed model is related to the LDA topic model framework but incorporates the idea of decomposing individual products (SKUs) into managerially relevant product attributes. It formulates the underlying topics as probability distributions over product attributes instead of individual products. We adopt the terminology “topic” from the natural language processing literature. A “topic” is similar to the underlying interest, need, objective, or goal that drives consumers’ shopping behaviors. Our model offers several advantages over LDA models in capturing consumer preferences in large store assortments. First, it is capable of making preference predictions for new products that are not in existing assortments, which is particularly important for retailers with frequently changing assortments. Second, unlike LDA models, which cannot incorporate the impact of time-varying marketing-mix variables, our model is able to do so because a marketing-mix variable can be treated as a product attribute in the model. Third, whereas LDA models ignore similarities among products sharing common attributes, our model explicitly captures such similarities. For example, in the context of fashion products, a consumer may have a strong preference for a particular brand or color and thus is likely to buy different clothes of the same brand or color repeatedly, which would be reflected in our attribute-based topic model. Fourth, the topics revealed by our model are clearly defined by managerially relevant product attributes and offer directly actionable insights for retail managers. In contrast, the meaning of an identified “topic” or “motivation” from an LDA model needs to be inferred from the commonality of products that are loaded on the same topic, and thus it is subject to interpretation and ambiguity.¹ Finally, our model can handle data sparsity better than LDA models, because its estimation relies on co-occurrence of product attributes in a shopping basket. In contrast, estimations of LDA models rely on co-occurrence of individual products in a shopping basket, which is less frequent than co-occurrence of product attributes.

The managerial focus of our research is to study the preference structures that drive consumers’ browsing and purchase activities for the entire store assortment of an online retailer. We thus construct a model of both browsing and purchase activities, in which topics for driving browsing and purchasing can be different, yet related. We apply the proposed model to data from an online deal marketplace for fashion products. Fashion retailers face unique challenges when it comes to assortment management. Consumers rarely buy the same item more than once, and their preferences tend to evolve over time. Certain fashion products have particularly short lifespans. These so-called “fads” need to be replaced frequently (e.g., Sproles 1981). Therefore, fashion retailers need predictive models that can be adapted to frequently changing store assortments. Our attribute-based topic model provides a solution to address this need by taking advantage of the fact that although consumers’ preferences for individual products are frequently changing, their preference structures at the product attribute level are more enduring. Many consumers have favorite brands and colors across product categories. The apparel categories that they favor also tend to be stable. For example, some women prefer pantsuits while others prefer power dresses, even though both categories are considered formal workwear.

In the context of deal-oriented fashion e-commerce sites, some questions are particularly relevant. For example, are consumers drawn by prestigious brands on deep discounts or low absolute prices regardless of the original prices and percentage discounts? Do consumers browse and purchase different products, and if so, what kind of products tend to draw consumers' interest in browsing, and what products tend to be converted into purchases? How does shopping experience influence consumers’ fashion preferences? How do consumers differ in their preference structures and evolution patterns? Our model will reveal easy-to-interpret and directly actionable “fashion topics of interest” that drive shoppers’ browsing and purchase activities and the time-varying patterns of their relevance to each individual. These insights can be utilized to make various merchandising decisions and product recommendations by online retailers or marketplace operators.

We make two main contributions in this study. First, we propose an ABMM model at the SKU level for inferring consumers’ preference structures in large and frequently changing store assortments, which overcomes several limitations of previous models for large assortments. We demonstrate that the proposed model outperforms benchmark models in its predictive power in both estimation and holdout data, and that it is particularly powerful in predicting preferences for new products not in existing assortments. Second, our study reveals novel insights on the underlying preferences driving consumers’ browsing and purchase activities in an online deal marketplace for fashion products and how they evolve with shopping experiences. The proposed model can be utilized to improve a variety of merchandising decisions. We illustrate its managerial value through an application of personalized product recommendations and demonstrate that recommendations based on our model substantially outperform those based on benchmark models.

Relevant Literature

Our research builds on several streams of prior literature. On the methodology front, our proposed model is related to previous developments in (1) assortment and SKU-level modeling and (2) topic models in machine learning developed for natural language processing, which have also been applied by marketing academics. In terms of the substantive area, it is related to the literature on online shopping and purchase behaviors. Next, we provide a brief review of each stream of research and outline how our study compares with the prior literature.

Assortment and SKU-Level Modeling

To deal with sales response or consumer demand/choice at the SKU level, models proposed in the marketing literature have essentially adopted two approaches. The first approach explicitly contains SKU-specific intercept terms to capture preference for a SKU (e.g., Guadagni and Little 1983). A key limitation of this approach is that the number of parameters to be estimated grows proportionally to the number of SKUs, and thus it cannot accommodate even a fairly small assortment well. The second approach is to define and parameterize the preference for a SKU as the sum of preferences for each of its product attributes. A seminal work taking this approach is Fader and Hardie (1996). They propose a multinomial logit model of product choice using a set of product attributes that parsimoniously characterizes a large number of SKUs in a category. Their model prevents explosion in the number of parameters. Moreover, it enables preference predictions for new products that are new combinations of product attributes in the model and provides valuable insights for identifying line extension opportunities.

Many studies have followed this general approach and extended it to incorporate other aspects of consumer decisions. For example, the model proposed by Ho and Chong (2003) captures the impact of consumers’ attribute-level experience and product-specific experience. Inman, Park, and Sinha (2008) develop a model that identifies product attributes driving switching versus repeat purchase behaviors, respectively. Rooderkerk, Van Heerde, and Bijmolt (2013) accommodate substitution patterns between SKUs by incorporating attribute-level similarity in their model. Sinha, Sahgal, and Mathur (2013) use an attribute-based approach to modeling sales of 3,000 SKUs of wine products and develop a decision support system based on it. Researchers have also utilized factor model structures in this line of pursuit. For example, Sinha et al. (2005) develop a factor-analytic choice model which captures a large number of SKUs in the soup category by a few latent factors. Singh, Hansen, and Gupta (2005) use a factor structure to capture cross-category correlations in attribute preferences and to deal with the scalability problem. In the operations management literature, research on assortment management has focused on developing heuristics that can solve discrete allocation optimization problems of sizeable assortments. Kök, Fisher, and Vaidyanathan (2008) provide a comprehensive review of operations management research on this topic.

Attribute-based models can be scaled up to handle a large number of products and to make preference predictions for new products, which is an inspiration for our ABMM model for store-level assortments. By adopting the mixed-membership modeling approach, our proposed model provides more flexibility in capturing preference structures as combinations of product attribute levels with varying degrees of relevance and heterogenous preferences among individual consumers.² In addition, compared with category-level models—which, by construction, focus on a given category and thus do not capture co-occurrences of purchases across product categories—our model incorporates interconnections in consumer preferences across categories (and other attributes), which can provide valuable insights for cross-selling, product recommendations, and assortment management at the store level.

LDA Topic Models

As mentioned previously, LDA topic models were originally developed for natural language processing in the computer science field. They are unsupervised Bayesian learning algorithms mainly used to extract latent “topics” by relating the words under a topic (for details, see Blei, Ng, and Jordan [2003]). There have been a growing number of applications of the LDA modeling framework in the marketing literature, primarily adopted to understand latent dimensions extracted from rich text data (e.g., Ansari, Li, and Zhang 2018; Dew, Ansari, and Li 2020; Liu and Toubia 2018; Tirunillai and Tellis 2014; Toubia et al. 2019). For example, Liu and Toubia (2018) study consumers’ content preferences in online search queries. They develop a hierarchical LDA model that allows topics in search queries to be semantically related to topics in search results.

LDA topic models have also been adopted to model purchase decisions in store-level assortments. Jacobs, Donkers, and Fok (2016) apply the LDA framework to study shopping baskets in an online grocery store and to identify the underlying “motivations” that drive the purchases. Their model outperforms several benchmark methods such as collaborative filtering and Dirichlet-multinomial mixtures in its predictive capability. Jacobs, Fok, and Donkers (2021) extend their previous model by including effects of shopping-trip-specific time indicators and incorporating temporal correlations in topic relevance using a correlated topic model with a vector autoregression formulation, which allows consumers’ preferences to change over time. They apply the extended model to study purchase decisions in an offline hardware store. Despite the improvement, their approaches share the limitations of LDA models when studying consumers’ preferences in frequently changing store assortments, as discussed previously. In this study, we propose a new mixed-membership topic model that overcomes these limitations. A prominent distinction of our proposed model with LDA models is that it formulates topics as distributions over product attributes and thus is capable of making preference predictions for new products that are not in the calibration data. In addition, we study the underlying “topics” driving both browsing and purchase activities in an online fashion retailing context. Our analyses reveal novel insights about connections and differences of preference structures driving these two types of shopping activities and how they evolve with consumers’ shopping experiences. Table W1 in Web Appendix A summarizes key features of the aforementioned studies that employ SKU-level models and compares ours with theirs.

Online Browsing and Purchase Behaviors

Our research is also related to the literature on online browsing and purchase behaviors. Prior research has extensively examined online shopping and purchase processes and/or decision outcomes. For example, Bucklin and Sismeiro (2003) study online users’ browsing behaviors on a website and find that a visitors’ propensity to continue browsing is affected by the depth of a given site visit and the number of repeat visits. Montgomery et al. (2004) focus on the navigation paths of online browsing behaviors and show that a user's past path information reflects their goals and is crucial in predicting a subsequent path. Moe (2006) examines the varying decision rules in online browsing and purchase stages within a given product category using attribute-based models of consideration set formation and choices, which identify the most relevant attribute(s) at each stage. Shi and Zhang (2014) investigate how online shoppers’ usage experiences with various decision aids influence the evolution of their store visit and purchase spending decisions over time.

In contrast, the substantive focus of our study is to identify the underlying product interests (which we call “topics”) that drive online browsing and purchase activities in large and frequently changing store assortments. Online shopping processes can be distinguished between the browsing and purchase stages. Moe (2003) characterizes web navigation sessions as knowledge-building, searching, browsing, and buying visits and shows that these types of visits differ in purchase likelihood and responsiveness to marketing messages. The online browsing and purchase stages can be linked conceptually to the broader framework of the purchase funnel as well as to the consideration set formation and purchase/choice stages, which have been modeled extensively in previous studies using scanner panel data (e.g., Gilbride and Allenby 2004; Hauser and Wernerfelt 1990; Roberts and Lattin 1991).³ We expect that consumers exhibit broader product interests in the browsing stage than in the purchase stage. Therefore, the number of product attribute levels driving browsing activities is likely to be higher than that for purchase activities.

Model Formulation and Estimation

Our model is built on the mixed-membership modeling framework (for a comprehensive overview, see Airoldi et al. [2014]), which also includes LDA models. The basic premise of the model is that consumers browse or purchase certain products in a shopping session driven by a mixture of latent topics of interest. Figure 1 illustrates the main components of our model. The observed data of a consumer in a given shopping session are the products browsed and those purchased, if any. Each product is transformed into specific values defined by a set of product attributes. Topics are formulated as probability distributions over those product attributes. The topics have different degrees of relevance in driving shopping activities of a given consumer. Therefore, a consumer's preference structure is represented by individual-specific topic relevance probability distributions over the topics (for browsing and purchase activities, respectively), in combination with the nature of identified topics. We allow topics relevant to a consumer's browsing and purchase activities to be different but also related to each other.

Figure 1.

Main Components of the Proposed Model.

Formally, let i denote a consumer, i ∈{1, …, I}, and j denote a product (i.e., a SKU) in the store assortment, j ∈{1, …, J}. A shopping session is a visit to the online store during which a consumer browses certain products and possibly purchases some or even all of them. For each consumer i, we observe S_i sessions in total. The set of products browsed by consumer i in session s, s = 1, …, S_i, is denoted by $w_{is}^{[b]}$ = { $w_{is 1}^{[b]}$ , …, $w_{{isN}_{is}^{[b]}}^{[b]}$ }, with each element being a browsed product and $N_{is}^{[b]}$ being the number of unique products browsed in session s. Similarly, the set of products purchased by consumer i in session s is denoted by $w_{is}^{[p]}$ = { $w_{is 1}^{[p]}$ , …, $w_{{isN}_{is}^{[p]}}^{[p]}$ }, with each element being a purchased product and $N_{is}^{[p]}$ being the number of unique products purchased in session s. If no purchases are made in the session, $N_{is}^{[p]}$ = 0 and $w_{is}^{[p]}$ = Ø. By definition, the purchased products in session s are a subset of the browsed products (i.e., $w_{is}^{[p]}$ ⊆ $w_{is}^{[b]}$ and $N_{is}^{[p]}$ ≤ $N_{is}^{[b]}$ ).

In the rest of this section, we first describe the product attribute space and how topics are defined on this space, followed by the topic relevance distributions and the link between topics driving browsing activities and those driving purchase activities. We then present our method of identifying the optimal number of topics in a model and estimation techniques.

Product Attribute Space

To capture consumers’ preferences over product attributes, we represent each product/SKU w_j as a set of attribute values. The specific attributes for fashion products in our model are described subsequently. These attributes are captured in a vector w_jt = (c_j, w_jt), where c_j = [c_j1, …, c_ja, …, c_jA] contains discrete attributes of product j such as category, brand, gender of the target user (which we call “gender” for simplicity hereinafter), and color, with c_a denoting a given attribute and A being the number of discrete attributes. Following terminology in the conjoint analysis tradition, each specific value of a given product attribute (such as red for “color”) is called an attribute level. The number of levels for attribute c_a is denoted by |c_a|. The subvector y_jt = [y_j1t, …, y_jdt, …, y_jDt] contains continuous attributes of product j, where d denotes a given attribute, D is the number of continuous attributes, and t is the time period. In the empirical application, we include the original price, depth of discount, and number of days listed in the store, the latter two of which can vary over time. By representing a product/SKU at a given time point as a combination of specific product attribute levels, our model provides a convenient way to incorporate the impact of time-varying marketing-mix variables in driving consumer preference.

Topics: Probability Distributions over Product Attributes

As stated previously, a novelty of our model is that topics are characterized by probability distributions over product attributes (instead of over individual products as in LDA models). Each topic m ∈ {1, 2, …, M} is characterized by a set of distributions involving A + D individual probability distributions: A probability mass functions each for a discrete product attribute, and D probability density functions each for a continuous product attribute. Within topic m, a discrete product attribute c_a is represented by a categorical distribution over its attribute levels, denoted by the |c_a|-dimension probability vector $ϕ_{m}^{c_{a}}$ . The set of probability vectors ϕ_m = [ $ϕ_{m}^{c_{1}}$ , …, $ϕ_{m}^{c_{A}}$ ] reflects how discrete product attributes load on topic m. For example, a topic that reflects a preference for men's blazers and overcoats by retailers Zara and H&M would place high probabilities on “male” for the “gender” attribute, on “Blazers and Overcoats” for the category attribute, and on “Zara” and “H&M” for the brand attribute.

Distributions over continuous product attributes under topic m are captured by probability density functions with specifications chosen based on the measurement of each attribute. A positive product attribute without upper limits, d₁, such as the original price and days listed in the store, is assumed to follow a Gamma distribution: $y_{m}^{d_{1}}$ ∼ Gamma ( $λ_{1 m}^{d_{1}}$ , $λ_{2 m}^{d_{1}}$ ). A continuous attribute with values in the range of [0, 1], d₂, such as percentage discount, is assumed to follow a Beta distribution: $y_{m}^{d_{2}}$ ∼ Beta ( $ξ_{1 m}^{d_{2}}$ , $ξ_{2 m}^{d_{2}}$ ). Both Gamma and Beta distributions are flexible in capturing distribution functions of a variety of shapes. Our framework can also easily accommodate other distribution functions of continuous product attributes.

How Topics Drive Shopping Activities: Topic Relevance Distribution Probabilities

Like in an LDA model, the browsing of a product j in a given session s by consumer i is assumed to be driven by a given topic, denoted by $z_{isj}^{[b]}$ . The selection of this topic follows a categorical distribution over the M possible topics, which is represented by the individual- and session-specific probability vector $θ_{is}^{[b]} = [θ_{is 1}^{[b]}, \dots, θ_{ism}^{[b]}, \dots, θ_{isM}^{[b]}]$ . While the browsing of a given product is driven by one topic, multiple topics can influence the browsing activities in a session, with each of them being selected from the same categorical distribution. Therefore, each element $θ_{ism}^{[b]}$ in $θ_{is}^{[b]}$ represents the relevance of topic m to consumer i in session s. $θ_{is}^{[b]}$ is assumed to follow a Dirichlet distribution. To incorporate correlations among topics, we model $θ_{is}^{[b]}$ by:

\begin{aligned} θ_{is}^{[b]} (α_{is}) = & (\frac{\exp (α_{is 1})}{\sum_{m^{'} = 1}^{M} \exp (α_{is m^{'}})}, \dots, \\ \frac{\exp (α_{ism})}{\sum_{m^{'} = 1}^{M} \exp (α_{is m^{'}})}, \dots, \\ \frac{\exp (α_{isM})}{\sum_{m^{'} = 1}^{M} \exp (α_{is m^{'}})}) \end{aligned},

(1)

where α_is = (α_is1, …, α_ism, …, α_isM) is a vector that reflects consumer i's inclination toward the topics in session s. This formulation is an extension of the correlated topic model, originally developed by Atchison and Shen (1980) and Blei and Lafferty (2007), and is also adopted by Jacobs, Fok, and Donkers (2021). A larger value of α_ism leads to a larger expected value of θ_ism, which indicates a higher probability of topic m in session s for consumer i. The complete specification of

θ_{is}^{[b]}

in our model is related to the hierarchical Dirichlet scaling process approach, which we use to identify the number of topics (to be described subsequently). We present the formulation for time-varying α_ism's in the next subsection.

Evolution of Topic Relevance Distribution Vectors

Prior research shows that online purchase behaviors evolve with consumers’ experiences with an online store (e.g., Shi and Zhang 2014). To capture how each consumer's preference structure may change over time, we model the topic distribution vector α_is as functions of time and consumers’ shopping experiences with the focal store. Specifically, each α_ism is modeled as:

α_{ism} = κ_{im} + β_{m}^{^{'}} X_{is} + \sum_{a = 1}^{A} \sum_{l = 1}^{| c_{a} |} f_{alm} F_{isal} + ϵ_{ism},

(2)

where κ_im is a consumer- and topic-specific intercept common across different browsing sessions by a consumer and captures unobserved heterogeneity due to factors unique to each individual. ɛ_ism is the error term, which is assumed to follow a Normal distribution with mean μ_m and covariance matrix Σ_m. X_is is a vector containing predictor variables specific to a session s of consumer i that may shift the relevance of topics at the consumer- and/or session-level.⁴ Quarterly dummy variables are included in X_is to capture seasonality in the preference for fashion products. We are particularly interested in whether and how a consumer's preference structure may evolve over time as they accumulate more shopping experience in the focal online store. Therefore, we include in X_is several covariates that represent a consumer's cumulative shopping experience prior to the current session, including the number of prior browsing sessions, which captures the number of total exposures to the focal store; the cumulative number of categories browsed previously and the cumulative number of categories purchased previously, which reflect the breadth of a consumer's exposure to the store's assortment; and the cumulative prior store spending, which captures the extent of a consumer's transactional relationship with the focal store. Estimation results of the parameter vector β_m will reveal how each aspect of store experience may drive consumers’ preference evolution in the store. Such insights can enhance online retailers’ ability to predict future preferences.

In addition, we incorporate potential variety-seeking or inertia behaviors in the model. Using the product attribute “brand” as an example, as a consumer becomes more familiar with a given brand, their interest may switch to other brands, which would lead browsing activities to exhibit variety seeking. In contrast, other consumers may gravitate toward the brand(s) that they become more familiar with, and their browsing activities would exhibit inertia. Our model allows both variety seeking and inertia with the following specification. Following Ho and Chong (2003), we operationalize familiarity with a certain attribute level in a session, F_isal, as the logarithm of the number of times consumer i had browsed attribute level l (e.g., brand B1) of the attribute c_a (e.g., brand) before session s. f_alm is the coefficient for F_isal. The behavior pattern revealed depends on the nature of a topic and the sign of the relevant coefficients f_alm's in Equation 2. Suppose that brand B1 has a high distribution probability in topic 1, then f_{brand1,topic1} > 0 indicates that familiarity with brand B1 would increase the relevance of topic 1, which suggests inertia behavior around brand 1; in contrast, f_{brand1,topic1} < 0 would reveal variety-seeking behavior regarding brand B1. In general, positive coefficients of “own attribute levels” indicate inertia, positive coefficients of “cross attribute levels” and negative coefficients of “own attribute levels” indicate variety seeking, where “own” and “cross” are in terms of attribute levels with high versus low loadings on a given topic. We incorporate variety seeking and inertia of all discrete product attributes in the model.⁵

A drawback of including these attribute-level familiarity variables in α_ism is that, as the numbers of product attributes and attribute levels increase, the model is likely to suffer from overfitting. We address the potential overfitting problem by adopting the variable selection and regularization method LASSO (Tibshirani 1996). This method automatically determines the relevant attributes in F_isal. Following common practice, we impose double exponential (Laplace) priors on f_alm to induce sparsity on the parameters (Gelman et al. 2013, p. 368).

We further incorporate correlations between topics at the session level by specifying that the consumer-specific vector κ_i = (κ_i1, …, κ_iM) follows a multivariate Normal distribution with mean μ_κ and covariance matrix Σ_κ.⁶ This formulation enables us to capture richer relationships between topics, and therefore between product attributes. The ability to identify such relationships would offer additional insight for designing cross-selling recommendations.

Likelihood Functions of Browsing Activities

Let $w_{isj}^{[b]}$ = [ $c_{isj}^{[b]}$ , $y_{isj}^{[b]}$ ] denote product j browsed by consumer i in session s. D₁ is the number of positive continuous attributes without upper limits, and D₂ = D − D₁ is the number of positive continuous attributes with values in the range of [0, 1]. Let d₁ ∈{1, …, D₁}, d₂ ∈{D₁ + 1, …, D}. Our model assumes the following data generating process for browsing activities by consumer i in session s:

Choose $θ_{is}^{[b]}$ ∼ Dirichlet(α_is).

For each of the $N_{is}^{[b]}$ browsed products, $w_{isj}^{[b]}$ :

Choose topic $z_{isj}^{[b]}$ ∼ Categorical_M( $θ_{is}^{[b]}$ ).

Choose each product attribute in $w_{isj}^{[b]}$ = [ $c_{isj}^{[b]}$ , $y_{isj}^{[b]}$ ] according to the probability distributions in $z_{isj}^{[b]}$ (which is simplified to z):

Choose level l of discrete attribute c_a, l ∼ Categorical_|ca|( $ϕ_{z}^{c_{a}}$ ), c_a ∈{c₁, …, c_A}.

Choose $y_{isj}^{[b], d_{1}}$ ∼ Gamma( $λ_{1 z}^{d_{1}}$ , $λ_{2 z}^{d_{1}}$ ), d₁ = 1, …, D₁.

Choose $y_{isj}^{[b], d_{2}}$ ∼ Beta( $ξ_{1 z}^{d_{2}}$ , $ξ_{2 z}^{d_{2}}$ ), d₂ = D₁ + 1, …, D.

The probability that topic m drives the browsing of

w_{isj}^{[b]}

is Pr(

z_{isj}^{[b]}

= m) =

θ_{ism}^{[b]}

. We account for all possible topic assignments to obtain the marginal likelihood that consumer i browses product j in session s, resulting in:

\begin{aligned} L (w_{isj}^{[b]} = (c_{isj}^{[b]} = c, y_{isj}^{[b]} = y) | {ϕ_{m}^{c_{a}}}_{m \in {1, \dots, M}}^{a \in {1, \dots, A}}, \\ {λ_{m}^{d_{1}}}_{m \in {1, \dots, M}}^{d_{1} \in {1, \dots, D_{1}}}, {ξ_{m}^{d_{2}}}_{m \in {1, \dots, M}}^{d_{2} \in {D_{1} + 1, \dots, D}}, θ_{is}^{[b]}) \\ = \sum_{m = 1}^{M} [\prod_{a = 1}^{A} ϕ_{m, c_{a}}^{c_{a}} \times L (y_{isj}^{[b]} = y | {λ_{m}^{d_{1}}}_{}^{d_{1} \in {1, \dots, D_{1}}}, \\ {ξ_{m}^{d_{2}}}_{}^{d_{2} \in {D_{1} + 1, \dots, D}}, z_{isj}^{[b]} = m)] θ_{ism}^{[b]}, \end{aligned}

(3)

where

λ_{m}^{d_{1}}

is the vector of Gamma distribution parameters for

y_{m}^{d_{1}}

, and

ξ_{m}^{d_{2}}

is the vector of Beta distribution parameters for

y_{m}^{d_{2}}

. The likelihood of observing the set of

N_{is}^{[b]}

products browsed in session s by consumer i is the multiplication of Equation 3 across these products.

Linking Browsing and Purchase Activities

Browsing a product may not lead to purchasing it, but a purchase must be preceded by browsing, which implies that the underlying preference structures driving browsing and purchasing are innately different in some ways yet also interconnected. To capture the similarities and differences in the preference structure that drives the two types of shopping activities, we adopt the hierarchically dual LDA formulation developed by Liu and Toubia (2018). Our application of this formulation involves two premises. First, purchase and browsing activities share the same set of potential topics but are subject to different mixtures of topic relevance probability distributions even within the same session. Second, the relevance of topics driving purchases can be reinforced or weakened by topics that are relevant to browsing activities in the same session.

Formally, $θ_{is}^{[p]}$ , the topic relevance distribution probabilities corresponding to purchase outcomes, are specified to be functions of $θ_{is}^{[b]}$ . We capture the mapping between the two types of shopping activities in a matrix R, in which each element r_mm′ indicates how the relevance of topic m for browsing affects that of topic m′ for purchases, with the following parameterization of a Dirichlet distribution (see Liu and Toubia 2018):

\begin{aligned} θ_{is}^{[p]} (R, θ_{is}^{[b]}) = (\frac{\exp ({R^{'}}_{1}, θ_{is}^{[b]})}{\sum_{m' = 1}^{M} \exp ({R^{'}}_{m^{'}}, θ_{is}^{[b]})} \dots, \\ \frac{\exp ({R^{'}}_{m}, θ_{is}^{[b]})}{\sum_{m' = 1}^{M} \exp ({R^{'}}_{m^{'}}, θ_{is}^{[b]})}, \dots, \frac{\exp ({R^{'}}_{M}, θ_{is}^{[b]})}{\sum_{m' = 1}^{M} \exp ({R^{'}}_{m^{'}}, θ_{is}^{[b]})}) . \end{aligned}

( 4)

We put a Normal prior on each element in R. The exponential of the multiplication between R_m (the mth column of R) and $θ_{is}^{[b]}$ is proportional to the expected relevance of topic m for purchases in the session. To illustrate how this formulation captures the relationship between browsing and purchases, suppose M = 2, then the relevance of topics for purchases is as follows:

(θ_{is 1}^{[p]}, θ_{is 2}^{[p]}) = (\frac{\exp (r_{11} θ_{is 1}^{[b]} + r_{21} θ_{is 2}^{[b]})}{\sum_{m = 1}^{2} \exp ({R^{'}}_{m} θ_{is}^{[b]})}, \frac{\exp (r_{12} θ_{is 1}^{[b]} + r_{22} θ_{is 2}^{[b]})}{\sum_{m = 1}^{2} \exp ({R^{'}}_{m} θ_{is}^{[b]})}),

(5)

where r₁₁ indicates the effect of the relevance of topic 1 for browsing on the relevance of topic 1 for purchases, and r₂₁ indicates the effect of the relevance of topic 2 for browsing on the relevance of topic 1 for purchases in the same session. A positive (negative) value of r_mm′ indicates that the relevance of topic m for browsing is positively (negatively) associated with the relevance of topic m′ for purchases in a consumer's preference structure. Note that our model allows covariates in Equation 2, which drives

θ_{is}^{[b]}

, to have direct effects on purchase decisions. These effects are represented mathematically through Equation 4. The covariates include previous store experience and product attribute familiarity variables, among others. Similarities and dissimilarities of products browsed and purchased in the same session are captured by the common set of topics driving the session, as reflected in

θ_{is}^{[b]}

and

θ_{is}^{[p]}

The data-generating process for each of the products purchased by consumer i in session s, denoted by $w_{isj}^{[p]}$ , is similar to that of a browsed product, except that the vector of topic relevance distribution probabilities $θ_{is}^{[p]}$ is generated according to Equation 4. The conditional likelihood of observing the set of products purchased in session s by consumer i, $w_{is}^{[p]}$ = { $w_{is 1}^{[p]}$ , …, $w_{{isN}_{is}^{[p]}}^{[p]}$ }, given the set of products browsed in session s by consumer i, $w_{is}^{[b]}$ = { $w_{is 1}^{[b]}$ , …, $w_{{isN}_{is}^{[b]}}^{[b]}$ }, is the multiplication of the likelihood functions across products purchased in the session:

\begin{aligned} L (w_{is}^{[p]} | w_{is}^{[b]}) = \prod_{j \in w_{is}^{[p]}}^{} L (w_{isj}^{[p]} = (c_{isj}^{[p]}, y_{isj}^{[p]}) | {ϕ_{m}^{c_{a}}}_{m \in {1, \dots, M}}^{a \in {1, \dots, A}}, \\ {λ_{m}^{d_{1}}}_{m \in {1, \dots, M}}^{d_{1} \in {1, \dots, D_{1}}}, {ξ_{m}^{}}_{m \in {1, \dots, M}}^{d_{2} \in {D_{1} + 1, \dots, D}}, θ_{is}^{[p]}) . \end{aligned}

(6)

The full likelihood of observing the set of products browsed (

w_{is}^{[b]}

) and purchased (

w_{is}^{[p]}

) in session s by consumer i is

L (w_{is}^{[b]}, w_{is}^{[p]}) = L (w_{is}^{[p]} | w_{is}^{[b]}) L (w_{is}^{[b]})

We summarize the key notations of the proposed model in Table 1. In addition, a graphical model representation of our model is presented in Web Appendix B.

Table 1.

Summary of Notations in the Proposed Model.

Notation	Description
i	Index for a consumer, I = total number of consumers
j	Index for a product (SKU), J = total number of products
s	Index for a shopping session, S_i = total number of sessions for consumer i
$w_{is}^{[b]}$	The set of products browsed by consumer i in session s
$N_{is}^{[b]}$	Number of unique products browsed by consumer i in session s
$w_{is}^{[p]}$	The set of products purchased by consumer i in session s
$N_{is}^{[p]}$	Number of unique products purchased by consumer i in session s
c _j	Vector of discrete attributes of product j, c_j = [c_j1, …, c_ja, …, c_jA]
c_a	Index for a discrete attribute, A = total number of discrete attributes
\|c_a\|	Number of attribute levels in discrete attribute c_a
$ϕ_{m}^{c_{a}}$	A probability vector representing the categorical distribution over attribute levels of discrete attribute c_a in topic m
y _jt	Vector of continuous attributes of product j at time t
d	Index for a continuous attribute, D = total number of continuous attributes
d₁	Index for a positive continuous attribute without an upper limit, D₁ = total number of positive continuous attributes without upper limits
d₂	Index for a continuous attribute with values in the range of [0,1], D₂ = D − D₁ is the total number of continuous attributes with values in the range of [0,1].
$λ_{1}^{d_{1}}$	Shape parameter of the Gamma distribution for d₁, a positive continuous attribute without an upper limit
$λ_{2}^{d_{1}}$	Rate parameter of the Gamma distribution for d₁, a positive continuous attribute without an upper limit
$ξ_{1}^{d_{2}}$	Shape parameter of the Beta distribution for d₂, a positive continuous attribute with values in the range of [0,1]
$ξ_{2}^{d_{2}}$	Rate parameter of the Beta distribution for d₂, a positive continuous attribute with values in the range of [0,1]
m	Index for a topic, M = total number of topics
$θ_{is}^{[b]}$	Vector of topic relevance distribution probabilities for consumer i's browsing activities in session s; $θ_{is}^{[b]} \sim Dirichlet (α_{is})$ .
α_is.	Vector of parameters that characterize the Dirichlet distribution of $θ_{is}^{[b]}$
κ_im	A consumer- and topic-specific intercept common across browsing sessions of a consumer
X _is	Vector of explanatory variables specific to consumer i in session s that may shift topic relevance through affecting α_is.
β _m	Vector of parameters for X_is, indicating how X_is. drives α_is
F _is	Vector of attribute familiarity variables of consumer i in session s
f _m	Vector of parameters for F_is, indicating how F_is. drives α_is
R	M × M matrix indicating how the relevance of topics for browsing affects that of topics for purchases
$θ_{is}^{[p]}$	Vector of topic relevance distribution probabilities for consumer i's purchase activities in session s
$z_{isj}^{[b]}$	The topic that drives the browsing of product _j by consumer i in session s
$z_{isj}^{[p]}$	The topic that drives the purchase of product j by consumer i in session s

Number of Topics

The number of topics (M) in mixed-membership models is a hyperparameter. It is typically prespecified to a certain value before the estimation of a model (e.g., Blei, Ng, and Jordan 2003), and the optimal number of topics is determined by comparing alternative models based on predictive performance measures such as deviance information criterion or Watanabe–Akaike information criterion (WAIC). In addition to being computationally expensive, this approach tends to favor models with a large number of topics, which would lead to overfitting of the data (Sharif-Razavian and Zollmann 2008). To overcome these problems, we adopt a Bayesian nonparametric method called the hierarchical Dirichlet scaling process (HDSP), which assumes an infinite number of topics a priori and identifies the number of relevant topics in a parsimonious way through the model estimation process (Kim and Oh 2017; Okui 2020; Paisley, Wang, and Blei 2011; Teh et al. 2006). Recently, Boughanmi and Ansari (2021) used this approach to infer the number of themes in a mixed-membership representation of music albums in the topic space. Details of our implementation of the HDSP method are provided in Web Appendix C.

Estimation

We use Markov chain Monte Carlo (MCMC) techniques to estimate the parameters in the model. Specifically, we use the no-U-turn sampler variant of Hamiltonian Monte Carlo (HMC) developed by Hoffman and Gelman (2014). This sampling method explores a high-dimensional posterior distribution exhibiting high autocorrelations over the MCMC draws much more efficiently than the Gibbs sampler and Metropolis algorithm (Gelman et al. 2013). The implementation is done in the software Stan (Stan Development Team 2018). We parallelize calculations of the log-likelihood across 12 processors using the reduce_sum function in Stan, which speeds up the sampling process of HMC significantly. Estimation of the proposed model takes 159.8 hours when using 12 (Intel Xeon W-2295) 3.0 GHz processors. Web Appendix D provides details of prior specifications in the proposed model. Synthetic data tests show that our estimation procedures recover the parameters well (details are provided in Web Appendix E). We include the Stan codes of model estimation in Web Appendix F.

Empirical Setting and Data

Institutional Background

We apply the proposed modeling framework to data from an online deal marketplace based in China that specializes in fashion products. It was launched in April 2013 as a marketplace for merchants to sell products at discounted prices for limited periods. Via the company's mobile app, consumers can view detailed information on available products, such as their picture, brand, product descriptions, the original price and sale price, and days listed on the app, in addition to placing orders. This marketplace operates exclusively through its mobile app and does not have a web presence. Online deal marketplaces are an ideal context for the purpose of our research because products sold in such venues are usually available for only a limited period, and thus the assortment is frequently changing.

Data Description

Our data set contains individual consumers’ browsing and purchase records with detailed timestamps of their activities and information on all products offered on the marketplace from August 2013 to July 2014, spanning a total of 53 calendar weeks. The products in our data were from the inventory management database, which only included products that were featured on the selling pages. Information on each product (SKU) included its category, brand name, color, target users (such as men or women⁷), original price,⁸ sale price, and days listed on the app by a given time point. We eliminated a small number of SKUs that were browsed fewer than ten times in the entire data period, which yielded a total of 7,557 unique SKUs for model estimation and follow-up analyses.

We define a browsing session as a period of continuous browsing activities separated by no more than one hour of inactivity (similar to Schweidel and Moe [2016]). We use data of the first 45 weeks for model estimation and data of the last eight weeks for holdout validations. We use the holdout data set to assess the predictive performance of the proposed method and compare it with several benchmark models. We keep consumers who had at least three browsing sessions in the estimation data set in order to identify consumer-specific parameters, which leads to 1,329 consumers who had a total of 18,744 browsing sessions, 2,426 of which resulted in purchases. Of those consumers, 309 of them appeared in the holdout period.

Table 2 summarizes the full data set and the two sub–data sets. Of the 1,672 unique SKUs in the holdout data set, 753 appeared in the estimation data, and 919 did not exist in the estimation data, which reflects the frequently changing nature of the focal online store's assortment. This enables us to examine the model's ability to predict preference for SKUs in the existing assortment as well as new products not in the estimation data. We use the following product attributes in the analyses: category, brand, gender of the target consumer (“gender” hereinafter), and color, which are discrete; and the original price, depth of discount, and days listed, which are continuous. The brand names are recoded into B1, B2, …, B24 for confidentiality reasons. The depth of discount is computed based on a product's sale price and original price and measured as a percentage of the latter. “Days listed” is the number of days by time t since a product first appeared on the app, and this information was visible to consumers. Table 3 provides more details of the product attributes used in our analyses. In terms of the number of SKUs, the largest category was Tees & Knits (16.1%), followed by Loafers & Other Casual Shoes (11.6%), Shirts & Blouses (11.6%), Jackets & Blazers (9.1%), and Pants (8.3%). A brand appeared in 12.1 categories on average, and 63.5% of the SKUs were products for women. As for product colors, the most common one in terms of percentage of SKUs was Black (21.6%), followed by Blue/Navy (19.1%), White/Off-White (12.4%), Red (8.3%), and Gray (8.1%).

Table 2.

Description of the Estimation and Holdout Data Sets.

Data Set	# Consumers	# SKUs	# Sessions	# Purchase Occasions
Full data	1,329	7,557	20,665	3,084
Estimation data	1,329	6,638	18,744	2,426
Holdout data	309	1,672	1,921	658

Table 3.

Description of Product Attributes.

Discrete Attributes	Number of Levels	Attribute Levels
Category	22	Tees & Knits, Shirts & Blouses, Dresses, Pants, Jackets & Blazers, Sweaters, Overcoats, Jeans, Skirts, Lounge & Sleepwear, Underwear & Lingerie, Shorts, Loafers & Other Casual Shoes, Dress Shoes/Pumps, Boots/Booties, Jewel, Scarves & Wraps, Belts, Other Accessories, Handbags, Backpacks, Other bags
Brand	24	Coded as B1, B2, …, B24
Gender	2	Female, Male
Color	12	Black, Purple, White/Off-White, Red, Green/Deep Green, Blue/Navy, Yellow/Gold/Orange, Gray, Brown/Khaki, Pink, Mixed Colors, Other Colors
C ontinuous Attributes	Mean	SD	25th Percentile	Median	75th Percentile
Original price (yuan)	952.36	906.17	439	639	1,099
Depth of discount (%)	.34	.22	.12	.32	.52
Days listed	7.33	6.15	3.46	5.95	9.51

Table 4 presents descriptive statistics of the browsing and purchase activities of the 1,329 consumers in the estimation data. We use the number of categories browsed, the number of categories purchased, the cumulative spending prior to the current session, and product-attribute-level familiarity variables to capture consumers’ prior shopping experience. The familiarity variables are omitted in Table 4 due to the large number of them. Coefficients of variation show that there were large variations across consumers, especially in their prior purchase activities and total spending.

Table 4.

Descriptive statistics of the Estimation Data.

Variable	Mean	Median	SD	Min	Max
Number of browsed products per session	3.99	2.00	5.77	1	153
Browsed ≥2 products in a session (yes = 1, no = 0)	.61	1.00	.49	0	1
Number of purchased products in a session with any purchases	1.34	1.00	1.02	1	12
Number of categories browsed per session	2.60	2.00	2.22	1	18
Number of categories purchased per session	1.20	1.00	.59	1	7
Made any purchases in a session (yes = 1, no = 0)	.10	0	.30	0	1
Bought ≥2 products in a session with any purchases (yes = 1, no = 0)	.18	0	.38	0	1
Number of browsing sessions per consumer	14.10	6.00	24.54	3	261
Number of purchase occasions per consumer	1.36	0	5.03	0	135
Cumulative number of categories previously browsed	11.34	12.00	6.75	0	22
Cumulative number of categories previously purchased	1.22	0	2.67	0	18
Cumulative store spending in the past (in 100 yuan)	12.04	0	56.94	0	971.29

Estimation Results

Model Comparisons

We compare our model with three benchmark models. Benchmark Model 1 assumes that topic relevance distribution vectors for browsing and purchases in a session are independent (i.e., without the linkage via the matrix R in the proposed model). In this model, each vector for a consumer is estimated separately as functions of attribute familiarities and store experience. Benchmark Model 2 treats topics with the LDA approach, in which each topic is represented as a probability distribution over individual SKUs in the assortment, and it includes the linkage between browsing and purchase vectors via the matrix R. This is essentially the model by Liu and Toubia (2018), which nests the model by Jacobs, Donkers, and Fok (2016). Benchmark Model 3 is the application of the LDA model by Jacobs, Fok, and Donkers (2021) to our context. It incorporates temporal correlations of topic relevance across sessions using a first-order vector autoregression formulation. This model allows topic relevance to change over time using a parsimonious method, but it does not provide insights on what drives a consumer's topic relevance evolution. To make a fair comparison, we use the HDSP approach for determining the number of topics for all models and incorporate our formulation of topic relevance distribution vector evolutions in Benchmark Models 1 and 2. The number of topics identified for the proposed model and Benchmark Models 1–3 is 14, 9, 15, and 15, respectively.

We use the WAIC for model comparisons. WAIC (Watanabe 2010) is a fully Bayesian measure of out-of-sample predictive performance that considers the entire posterior distribution instead of conditioning on a point estimate. It is particularly desirable in our context, which involves models with hierarchical and mixture structures, and thus the number of parameters increases with the sample size (Gelman, Hwang, and Vehtari 2014). A lower WAIC indicates better predictive performance of a model. We compute and compare WAICs across models in the estimation and holdout data sets, respectively, and further separate out WAICs for existing versus new products in the holdout data.

By definition, new products in the holdout data did not exist in the estimation data. We can directly compute WAICs on all products in the holdout data for the proposed model and Benchmark Model 1, because they formulate topics over product attributes and thus can handle new SKUs. To compute WAICs in the holdout data for Benchmark Models 2 and 3, we assume that the probability of observing a new SKU equals the smallest probability among existing SKUs in the corresponding topic assignment, because the probability of observing a new SKU given a topic assignment must be smaller than the probability of observing any existing SKU given the same topic assignment.

Tables 5 and 6 present WAICs of the proposed model and benchmark models for browsing and purchase activities in the estimation and holdout data sets, respectively. Since Benchmark Models 2 and 3 are not designed to make preference predictions for new products, we expect their performance to be worse for new products in the holdout data. As shown in the tables, our model in general performs better than the three benchmark models in both the estimation and holdout data sets based on WAICs for the entire data set (browsing and purchases). In terms of the ability to predict browsing or purchase activities separately, the proposed model outperforms the benchmark models for both activities in the estimation data and also does so in the holdout data, with the exception that WAIC of Benchmark Model 1 is slightly lower than the proposed model for browsing activities. It is worth noting that the two attribute-based models (the proposed one and Benchmark Model 1) perform substantially better in predicting browsing and purchases in the holdout data than Benchmark Model 2, which is an LDA model, because of the former's ability to predict preferences for new products. In addition, compared with Benchmark Models 1 and 3, which do not account for the interconnections of topic relevance between browsing and purchases, the proposed model performs better in predicting purchase activities in both estimation and holdout data sets.⁹

Table 5.

Model Comparisons: WAICs in the Estimation Data Set.

Shopping Activity	Proposed Model	Benchmark Model 1	Benchmark Model 2	Benchmark Model 3
Browsing	1,122,668.46	1,216,486.93	1,144,706.21	1,216,250.04
Purchases	33,057.09	168,196.66	36,049.51	102,964.93
Browsing and purchases	1,155,758.76	1,384,683.59	1,181,658.94	1,319,214.97

Table 6.

Model Comparisons: WAICs in the Holdout Data Set.

		Proposed Model	Benchmark Model 1	Benchmark Model 2	Benchmark Model 3
Browsing	Existing products	84,282.90	81,593.25	68,536.53	65,203.41
	New products	119,855.59	117,859.15	396,792.70	193,002.00
	All products	204,138.49	199,452.40	465,329.23	258,205.41
Purchases	Existing products	7,041.89	9,327.60	8,547.46	5,957.53
	New products	10,936.99	13,374.42	22,998.43	12,822.62
	All products	17,978.88	22,702.02	31,545.89	18,780.14
Browsing and purchases	Existing products	91,324.81	90,920.85	77,083.99	71,160.93
	New products	130,792.58	131,233.57	419,791.10	205,824.62
	All products	222,117.39	222,154.42	496,875.09	276,985.55

The model comparisons demonstrate the importance of incorporating the interconnections of topic relevance for browsing and purchase activities in our model, which enables it to do a much better job of predicting preferences for purchases than Benchmark Model 1. It enhances the model's ability to address the challenge brought about by sparsity of purchase occasions in the data, because it pools information in both purchase and browsing activities, the latter of which occurs more frequently. Compared with Benchmark Models 2 and 3, which use the LDA approach to representing topics, our attribute-based topic model can predict preferences for new products, which is a major advantage for retailers with frequently changing assortments.

Product attributes in the proposed model need to be defined at a high level so that they are applicable to all merchandise in the store. A potential modification of it is to apply the model at merchandise category level and to incorporate category-specific attributes. We have tested such category-level models in three categories: Jackets & Blazers, Pants, and Shirts & Blouses. Comparisons of WAICs in the estimation and holdout data sets indicate that these category-level models perform substantially worse than the proposed store-level model in predicting browsing and purchase activities in each category (details of the models and comparison results are reported in Web Appendix G). The main reason is that co-occurrences of product attributes in a shopping basket are much sparser in category-level data than in store-level data, which severely hinders the reliability of parameter estimates and predictive power of the category-level topic models. In addition, the proposed store-level model incorporates associations in preference structures across categories, which cannot be captured by the category-level models. In conclusion, we do not recommend implementing our proposed topic model to category-level data, especially when co-occurrences of product attributes are sparse.

We report estimation results of key parameters of the proposed model in Web Appendix H and focus on describing the substantive findings in the following subsections.

Topics

Our model identifies 14 topics based on the HDSP approach. In each topic, the distribution over a discrete attribute is represented by a probability vector over its attribute levels, the original price and days listed in the store are modeled using Gamma distributions, and the depth of discount (%) is modeled using Beta distributions. Figure 2 presents a heat map of the topics characterized by their estimated probability distributions of discrete product attributes (category, brand, gender, and color). Each row represents a topic, and each column represents a specific product attribute level, with heat colors indicating the expected probability of a given attribute level. The numbers in the last three columns represent the mean of days listed (in days), the original price (in 100 yuan), and depth of discount (%), respectively, of the distribution represented by each topic. Figure 3 presents the distribution plots of the three continuous attributes.

Figure 2.

Heat Map of Topics.

Figure 3.

Distributions of Continuous Attributes in Each Topic.

Figure 2 shows varied patterns in terms of how specific product attributes are related to the identified topics. Some topics are characterized by a single attribute level on most product attributes, such as Topic 2. Nonetheless, most other topics are characterized by attributes spanning across multiple levels of all physical attributes. Figure 3 shows that the distributions of the continuous attributes take on quite different shapes. The distribution of days listed is fairly similar across topics. All of them are highly right-skewed, which indicates that consumers prefer products that are newly listed on the deal marketplace in general. The distributions of the original price and depth of discounts show a high degree of variation in both shapes and means across topics, which reflects a wide spectrum of price preferences in this deal marketplace. The original price of a fashion product reflects its brand image and prestige (Groth and McDaniel 1993; Lee, Chen, and Wang 2015). While a couple of topics represent preferences for low original prices and high depth of discounts (e.g., Topics 10 and 12), others indicate the opposite preferences for high original prices and relatively low depth of discount (e.g., Topics 6 and 11). Some other topics capture preferences for low original prices and low discounts (e.g., Topics 5 and 13) or for high original prices and high depths of discount (e.g., Topic 1). As we illustrate subsequently, these findings offer the basis to segment consumers into different profiles of price preferences, which can assist a variety of merchandising decisions for a deal marketplace. To assist interpretations of the topics, in Table 7 we provide more details of the physical product attributes that characterize each topic.

Table 7.

Descriptions of the Identified Topics.

Topic	Summary	Detailed Description
1	Women's designer clothes in conservative colors	Women's jackets and blazers, overcoats, sweaters, and shirts and blouses in black, blue/navy, and gray colors, premium brands (B2, B4, B5, B13)
2	Women's bags	Women's backpacks and handbags in multiple colors, value brand B6
3	Women's casual clothes in dark colors	Women's tees and knits, shirts and blouses, and pants, in black and gray colors, premium and mid-level brands (B5, B12)
4	Women's basic warm-weather clothes	Women's tees and knits, shirts and blouses, and dresses in white, blue/navy, and black colors, value to mid-level brands (B18, B20, B21)
5	Men's designer clothes in dark colors	Men's jackets/blazers, shirts and blouses, tees and knits, and pants in blue/navy and black colors, premium brands (B10, B14, B22)
6	Men's relaxed outfits	Men's jackets/blazers, jeans, and loafers and other casual shoes, in black and blue/navy colors, mostly premium brands (B9, B10, B19)
7	Casual shoes for men and women	Men's and women's loafers and other casual shoes in multiple colors (black, blue/navy, green, red), value brand B1
8	Women's casual tops in modern colors	Women's tees and knits and sweaters in black, pink, mixed colors, and gray, mid-level and premium brands (B7, B18)
9	Men's casual clothes in basic colors	Men's tees and knits, shirts and blouses, and pants in black, grey, blue/navy, and white colors, mid-level and premium brands (B5, B8, B10)
10	Women's outerwear of a given brand	Women's overcoats, handbags, and scarves and wraps, in black, blue/navy, and brown/khaki colors, mid-level brand B7
11	Women's clothes of two sister brands	Women's overcoats, jackets and blazers, and dresses, in multiple colors (black, white, yellow, blue/navy), premium brand B18 and value brand B21, which belong to the same parent company
12	Women's tees & knits and inexpensive accessories	Women's fashion jewelry, tees and knits, handbags, scarves/wraps, in black and white colors, mostly mid-level brand B7
13	Dark-colored street fashion for men and women	Men's and women's loafers and other casual shoes, and tees and knits, in blue/navy, black, and gray colors, value to premium brands (B1, B14, B23)
14	Women's high-end accessories	Women's jewelry, handbags, and underwear and lingerie, in yellow/gold, pink, and black colors, mostly two premium brands (B16, B19)

Relations of Topic Relevance for Browsing Versus Purchase Activities

The interconnections of topic relevance for browsing and purchases are captured by the matrix R in our model (Equation 4). Based on the model estimation results, we compute each topic's average relevance probability across all shopping sessions in the data, for browsing and purchases, respectively. The key results are illustrated in Figure 4. In the middle of the figure is a color-coded graph of posterior means of elements in the matrix R, in which an element r_mm′ indicates how the relevance of topic m for browsing activities is associated with the relevance of topic m′ for purchase activities (numerical estimation results are reported in Web Appendix H, Table W8). Most diagonal elements (10 out of 14) of the matrix R are positive, indicating that each of these topics for browsing reinforces purchase activities. This means that many products being browsed also ended up being purchased. Nonetheless, four diagonal elements (Topics 2, 7, 10, 11) are negative. In addition, there is sufficient positive mass on the off-diagonal elements. The negative diagonal elements and positive off-diagonal elements in the matrix R indicate a shift in consumers’ preferences at the browsing and purchase stages.

Figure 4.

Relations of Topic Relevance for Browsing Versus Purchase Activities.

We further examine the similarities and differences of topics driving these two stages. In the top and bottom panels of Figure 4, the length of each bar represents the average relevance probability of each topic for browsing and purchases, respectively. As the figure shows, some topics play a prominent role in driving both browsing and purchase activities, such as Topic 4, which is characterized by multiple attribute levels within a discrete attribute and low original prices and lower depth of discounts. Nonetheless, browsing and purchases are not necessarily driven by the same topics. For example, Topic 1, which is characterized by high original prices and deep discounts, is more relevant in driving browsing than purchases across all consumers. In contrast, Topic 3, which is characterized by lower original prices and deeper discounts, is more relevant in driving purchases than browsing across all consumers.

To compare the breadth of attribute levels between browsing and purchases, we compute normalized entropies¹⁰ for each discrete product attribute based on the weighted average probabilities across attribute levels (weighted by topic distribution probabilities for a given session and consumer). For each of the continuous attributes, we compute the weighted mean and standard deviation weighted by the corresponding topic distribution probabilities. Table 8 presents the results. In general, the entropies for browsing are larger than those for purchases, indicating that consumers tend to browse a broader set of products than what they end up purchasing. The weighted standard deviation of the original price for browsing is also larger than that for purchases, which speaks to the same phenomenon. There is not much difference in the weighted standard deviations of days listed and depth of discount between the two stages. Interestingly, the weighted means of original prices and depth of discounts of topics driving browsing are substantially higher than those driving purchases, suggesting that, in general, consumers tend to be attracted by products of premium positioning and/or deep discounts on the deal marketplace in their initial search processes but gravitate toward products of lower tiers of original prices and modest depth of discounts for purchases.

Table 8.

Entropy of Distribution Probabilities of Discrete Attributes and Weighted Mean and Standard Deviation of Continuous Attributes.

	Normalized Entropy				Weighted Mean/Standard Deviation
Shopping Activity	Category	Brand	Color	Gender	Days Listed (Day)	Original Price (100 Yuan)	Depth of Discount (%)
Browsing	.88	.95	.91	.86	.75/.94	9.91/4.99	45.46/13.06
Purchases	.83	.92	.88	.89	.92/.93	3.00/3.72	21.30/13.46

Evolution of Topic Relevance

The evolution of a consumer's preference structure is captured by the effects of time-varying covariates in Equation 2 (β_m′). We report the parameter estimates in Web Appendix H and highlight the substantive findings next.

The role of consumers’ prior shopping experience

Our model shows how the relevance of each topic is associated with a consumer's prior shopping experience at the focal deal marketplace. For example, as consumers browse or purchase more categories from the marketplace, Topic 6 (characterized by high original prices and low depth of discounts) becomes more relevant, while Topic 10 (characterized by low original prices and high depth of discounts) becomes less relevant. This suggests that consumers may gravitate toward more expensive products as they have been exposed to more categories, which provides another reason for online fashion retailers to encourage consumers to browse more categories in their stores early on.

Effects of product attribute familiarities

Estimation results show that familiarity with attribute levels also drives preference evolution. The effects of a consumer's prior exposure to different product attribute levels on the topic relevance reflect variety seeking, inertia, or loyalty in consumers’ browsing or purchase behaviors. We use the variable selection method LASSO to determine the relevant product attribute familiarity variables for each topic. Web Appendix I reports detailed estimation results. We find that variety seeking was the dominant pattern in both browsing and purchase activities, especially with regard to category and color. The tendency to switch away from prior exposure to the color black is particularly strong (as indicated by negative effects of familiarity with this color on relevance of topics with high loadings of the black color) in both browsing and purchases. Nonetheless, consumers showed loyalty to several brands (brands B1, B7, B12, and B18), as familiarity with these brands reinforces the relevance of topics with high loadings on them. The estimation results also reveal cross-category browsing patterns across sessions. For example, as consumers gained more familiarity with the category Shorts, they tended to become more interested in browsing Jackets & Blazers and Overcoats.

Seasonality

The quarter dummy variables in Equation 2 capture potential seasonality in the preference for products in the assortment. The results are consistent with the expected seasonality of fashion products. For example, Topic 10, which is characterized by women's black overcoats, is most relevant in Quarter 4 (October–December), while Topic 12, which is characterized by women's black and white/off-white tees and knits, is most relevant in Quarter 3 (July–September).

Model-Based Consumer Profiles According to Price Preference

Understanding consumers’ price preferences is particularly important for deal-oriented online retailers or marketplaces. Some consumers may be attracted by prestigious brands (which are of high original prices) on deep discounts, while others may be attracted by low absolute prices regardless of whether it is due to low original prices and/or deep discounts. How important is the brand prestige versus discounts? How do consumer preferences for the original price and depth of discounts vary in the browsing vs. purchase stages? Our model incorporates the original price and depth of discount as continuous product attributes. It reveals consumers’ preferences for the two distinct dimensions of prices in browsing and purchases, respectively, which enable us to answer these questions by constructing a consumer profile according to their price preferences.

We profile consumers into four types based on their preferences for the original price and depth of discount. We first compute the average topic distribution vector across shopping sessions for each consumer, and then compute averages of original price and depth of discount weighted by their topic distribution probabilities for browsing and purchases, respectively. Next, we classify each consumer into “high” or “low” based on whether their average original price or depth of discount is above or below the corresponding average value across all consumers. This leads to four profiles of consumers based on their price preference. Table 9 summarizes the profiles and presents the percentage of consumers belonging to each profile for browsing and purchases, respectively.

Table 9.

Profiles of Consumers Based on Their Price Preference.

Characteristics	Profile	% of Consumers
Characteristics	Profile	Browsing	Purchases
High original price, low depth of discount	Prestige driven shoppers	15.34	11.88
High original price, high depth of discount	Smart shoppers (who care about prestige and savings)	27.46	28.66
Low original price, low depth of discount	Value oriented shoppers	35.43	39.8
Low original price, high depth of discount	Die-hard bargain shoppers	21.74	19.63

Table 9 shows that the majority of consumers on the focal deal marketplace belong to the middle two profiles in both browsing and purchases. The largest group is “value-oriented shoppers” who prefer products of low original prices but are not obsessed with high percentage discounts. The second largest group is “smart shoppers” who prefer prestigious brands with deep discounts. “Prestige driven shoppers” are the smallest group. Interestingly, more consumers belong to this group in their browsing behaviors than in purchases. Our analysis reveals that most consumers belong to the same profile for both browsing and purchases. Nonetheless, some customers show different profiles in their browsing and purchase activities. For example, 3.91% of consumers are value-oriented shoppers for browsing but die-hard bargain shoppers for purchases, while 5.64% of consumers are die-hard bargain shoppers for browsing but value-oriented shoppers for purchases. The individual differences provide valuable information for personalizing product offerings and communication messages based on each consumer's price preference for browsing versus purchases.

Our model provides a rich set of insights that can be used to assist or improve a variety of merchandising decisions, such as assortment planning, cross-selling, and designing personalized offerings. In the next section, we illustrate its managerial value through an application of personalized product recommendations.

Application: Personalized Product Recommendations

Personalized product recommendations have been widely used in e-commerce practices. A general challenge is the “cold start” problem of making recommendations for new products that do not exist in the model calibration data set/existing assortment. As discussed previously, LDA models of store assortments cannot make preference predictions for new products and thus are not designed to address this problem. Conventional collaborative filters rely on correlations of products purchased across users. They cannot address the problem either, because products not in the existing assortment by definition could not have been browsed or purchased by anyone previously. Solving the cold-start problem regarding new products is particularly relevant for retailers managing frequently changing assortments. For example, in our empirical application here, 55% of SKUs in the holdout data set were new products that did not exist in the calibration data set. Our model provides the basis for a solution to address this challenge.

A personalized product recommendation system identifies a certain number of products that are predicted to maximize a consumer's likelihood of the desired outcomes. In the empirical application, we set the goal to find a set of recommended products that maximizes a consumer's likelihood of browsing and purchasing. When applying our proposed model to derive the product recommendations, the basic idea is to find the optimal set of product attributes from combinations of available attributes in the assortment. Using the recommendation set of one product as an example, the said problem translates to maximizing the following likelihood over product attribute vectors: $ma x_{w_{j} = (c_{j}, y_{j})}$ L( $w_{j}^{[b]}, w_{j}^{[p]}$ ), W_j ∈ W, where W is the entire assortment of products which could include new products not existing in the model calibration data set. The search space for product attributes is constrained by combinations of product attributes used in the predictive model.

Solving this optimization problem involves two challenges. First, some product attributes take discrete values, whereas other product attributes take continuous values, which prevents us from adopting gradient-based optimization methods to maximize the objective function. Second, exploring the large combinatorial space of discrete product attributes for the exact solution that maximizes the likelihood using conventional methods such as grid or random search is a formidable task in terms of computing time and resources. To address these two challenges, we adopt the Bayesian optimization (BO) method to find optimal product attributes. Details of our application of the BO technique are presented in Web Appendix J. It takes 4.4 hours to complete the BO computation based on the proposed model for all consumers in the sample when using 12 (Intel Xeon W-2295) 3.0 GHz processors. In real-world applications, the predictive model only needs to be estimated once in a while. The BO derivation of product recommendations can be done even faster by utilizing more processors for parallel computing.

We compare the performance of personalized product recommendations based on our model with those based on the three benchmark models described previously. For each method, we create consumer-specific recommendation sets of products that are most likely to be browsed and purchased and then compare them with a consumer's actual browsed and purchased products in the holdout data set. As a reminder, Benchmark Model 1 is an attribute-based topic model like ours except it does not consider interconnections of the topic distribution probabilities for browsing and purchases. We apply the same BO method to derive the recommended products for this model. Benchmark Models 2 and 3 are LDA models with topics formulated over individual products. For these two models, we first compute a consumer's likelihood of browsing and purchasing each product in the calibration data directly from the model and then identify the top-ranking products based on the predicted likelihood.

A recommendation set of size V contains the V highest-ranked products for a consumer. We vary the size V to allow different contexts of personalization applications (Jacobs, Donkers, and Fok 2016). In the empirical application, we examine V ranging from one to ten, because the number of recommended products in a single push hardly goes beyond the upper limit of ten (Zolaktaf, Babanezhad, and Pottinger 2018). For example, as of January 2022, jcrew.com (“Customers Also Love”) and gap.com (“Customers Also Liked”) recommend five products each time on their websites.

We assess the quality of a recommendation set by matching its content against products actually browsed and purchased by a consumer in the holdout data. Denote these products by h_i for consumer i, and the number of unique products in h_i by u_i. The vector g_i contains the V recommended products for consumer i in their ranking order, with the first element, g_i1, being the optimal product with the highest predicted likelihood of browsing and purchase, L( $w_{j}^{[b]}$ , $w_{j}^{[p]}$ ), based on our proposed model and the benchmark models, respectively.

The quality of a recommendation set of size V can be measured by the number of products in the recommended set g_i that overlap with the set h_i: $Σ_{v = 1}^{V} I [g_{iv} \in h_{i}]$ , where I[.] is the indicator function that equals 1 if the argument inside the bracket is true, and 0 otherwise. This number should be calculated relative to the maximum number of possible hits, which is min(V, u_i), to ensure that it is comparable across recommendation sets of different sizes (Jacobs, Donkers, and Fok 2016). Thus, the hit rate for consumer i is defined as $Σ_{v = 1}^{V} I [g_{iv} \in h_{i}] / min (V, u_{i})$ . To further incorporate the notion that positions in a recommendation list affect consumer's reaction to the recommended products (Smith and Brynjolfsson 2001; Xu and Kim 2008), we also compute the ranking-weighted hit rate developed by Jacobs, Donkers, and Fok (2016). It is computed as $[Σ_{v = 1}^{V} I [g_{iv} \in h_{i}] \times w (v, V)] / \sum_{v = 1}^{\min (V, u_{i})} w (v, V)$ , where the weight for the vth-ranked recommended product is w(v, V) = 1 − (v − 1)/V.

In addition, we compare a frequently used performance measure in the machine learning literature, the F-measure, which combines precision and recall (Adomavicius and Tuzhilin 2005). Precision is the proportion of products in the recommendation list that match with products actually browsed/purchased in the holdout data, and recall is the proportion of products actually browsed/purchased in the holdout data that appear in the recommendation list. The precision and recall measures correct for the fact that recommendation set sizes may not be the same. Since these two measures of a given algorithm do not usually go in the same direction, the overall performance is often captured by the F-measure, a harmonic mean of precision and recall (Adomavicius and Tuzhilin 2005): [(recall⁻¹ + precision⁻¹)/2]⁻¹, with a higher value indicating a better overall performance.

We present the average hit rate, ranking-weighted hit rate, and F-measure across all consumers for product recommendations based on our proposed model and the three benchmark models, for browsing and purchases, respectively, in Figures 5 and 6. We further separate out the performance for existing products (those that were present in the estimation data) and new products (those that only existed in the holdout data). Note that a given recommendation set may contain both existing and new products, but we separately tally and report the performance of the recommendation sets by these two groups of products. The two LDA models, Benchmarks 2 and 3, cannot make preference predictions for new products. Therefore, a priori, we expect their performance to be poor on all new product related measures. Moreover, as Figure 5 shows, even for existing products, our model and Benchmark 1, both of which are attribute-based topic models, yield substantially better performance in recommending products for browsing than the two LDA models. Moreover, recommendations based on our model outperform those by Benchmark 1 for browsing existing and new products in general.¹¹

Figure 5.

Predictive Performance for Browsing.

Figure 6.

Predictive Performance for Purchases.

The advantage of our model over the benchmark models is particularly strong when it comes to making recommendations for purchases, as shown by Figure 6. Although Benchmark Model 1 has the ability to make preference predictions for new products in theory, it shows poor performance in our empirical application. It is unable to predict purchases of any new products in all recommendation sets examined and of existing products in recommendation set sizes between 1–7. The reason is that Benchmark 1 treats the topic distribution vectors for browsing and purchases in a session independently. Although the flexibility allows Benchmark 1 to achieve a slightly better WAIC in predicting browsing in the holdout data than the proposed model (see Table 6), the sparsity of purchase data makes it ill-fitted to make predictions of purchases. In contrast, our model incorporates interconnections between the topic relevance for browsing and purchases, which enables it to enhance predictions for purchases by drawing on information in the browsing data. The current application further demonstrates the importance of incorporating interconnections of topic relevance in browsing and purchases and points to an opportunity to deal with the sparsity problem of purchase data by utilizing online browsing data. As to recommendations based on the two LDA models, Benchmark 2 is unable to hit any purchased products in the holdout data; Benchmark 3, which employs an autoregressive formulation of temporal correlations of topic relevance, does a decent job (yet still worse than our model) of recommending existing products for purchases, but it is unable to recommend any new products.

In addition to improved WAICs of purchases in the holdout data (as shown in Table 6), the current application shows that product recommendations based on our model offer a much closer fit with consumers’ preferences and would lead to substantially better purchase outcomes than those based on all three benchmark models. Since the product recommendation application uses only data in the calibration period to predict consumer preferences in the holdout period, the comparisons in this section are a better reflection of the practical value of our proposed model than WAICs.

Summary and Discussion

We have developed an attribute-based mixed-membership topic model for discovering online consumers’ preference structures in large and frequently changing store assortments from their browsing and purchase activities. The proposed model overcomes several limitations of LDA topic models for handling store assortments. A main advantage of our model is that it is capable of making preference predictions for new products not in the existing assortment and, thus, is particularly useful to retailers with frequently changing store assortments. In addition, our model can incorporate the influence of marketing-mix variables and other time-varying product attributes, which is an important improvement for assisting merchandising decisions. We demonstrate the managerial value of our model through an application of personalized product recommendations, and our model shows particularly strong performance over benchmark models in predicting purchases of new products in the holdout data. It offers a reasonable tool to address the cold-start problem of new products for personalized product recommendations, which is an important issue for retailers with frequently changing assortments.

The managerial focus of our research is to study consumer preferences in browsing and purchase activities in online deal stores/marketplaces of fashion products. Our empirical analyses reveal many novel insights of shopping behaviors in such venues. For example, we find that consumers’ preference structures tend to shift from the browsing to the purchase stage. In general, consumers tend to be drawn to products of premium positioning and/or deep discounts in the browsing stage, while their purchases tend to gravitate toward products of lower tiers of original prices and modest depths of discounts. Nonetheless, there are large variations across consumers. Drawing on the model estimation results, we profile consumers in the focal deal marketplace based on their price preferences, which provides valuable insights for modifying assortment compositions and designing personalized merchandising offerings.

The proposed model offers an effective tool for retail managers to discover underlying preference structures that could span seemingly dissimilar or unrelated product attributes without imposing a priori structures. In addition to its advantage in assisting personalized product recommendations, insights revealed by the model can be used to improve a variety of other merchandising decisions. For example, the revealed correlation patterns among product attributes are valuable input for creating new styles by recombining attributes from existing styles and modifying them in a way that is most likely to suit a consumer's taste. They can also be utilized by personal styling service providers to curate personalized product combinations that cater to each individual's preference. In addition, they can be very useful for retailers to devise cross-selling offerings. Furthermore, our model estimation results reveal detailed depictions of individual-specific preference structures and how they may evolve over time with consumers’ shopping experience in the store. For example, we find that certain topics became more prominent, while others became less prominent, in driving a consumer's shopping activities, as they had been exposed to more product categories in the store through either browsing or purchases. Such insights can help retail managers refine personalized targeting strategies based on a shopper's prior activities in their store.

Although our specific model formulation is developed to study browsing and purchase activities in an online store of fashion products, the modeling framework can be easily modified for other retail contexts. For example, offline retailers can simplify the model to focus on purchase outcomes, in which case the model does not need the hierarchical structure component. Retailers selling other types of merchandise can adjust the specific product attributes and their levels used in the model to fit the nature of their businesses.

Our study has several limitations that provide opportunities for future research. First, consumers’ preferences for fashion products can be influenced by social sources, such as social networks, news outlets, and fashion magazines. While our model captures the result of social influence as part of a consumer's preference structure, it cannot separate out the net effect of such influence due to a lack of data on these sources. If data on social influence sources are available, they can be incorporated into the topic relevance distribution vector of our model. We leave this as an important direction for future research. Second, although our model can make preference predictions for new products that did not exist in the estimation data/existing assortment, it requires that those new products be sufficiently mapped to the product attribute levels included in the calibration model. Therefore, the model is not well suited for predicting preference of products with drastically new attributes that are hard to relate to previous products or categories with frequent introductions of new product attributes. Third, while showing strong performance in addressing the cold-start problem of new products for personalized recommendations, our model cannot make reliable preference predictions of new customers/users. An important area of research endeavor is to develop more advanced models and algorithms that can simultaneously solve the “cold start” problems of new products and new users. Finally, our model is built on predetermined product attributes that are clearly defined and directly actionable by retail managers. Future research can expand insights from our model by using machine/deep learning techniques to identify unexpected product features that drive consumer preferences. This approach can be particularly useful for designing fashion products.

To conclude, we hope our study will stimulate more research on inferring consumers’ preference structures from complex assortment data and on advancing understanding of browsing and purchase behaviors in diverse e-commerce environments.

Supplemental Material

sj-pdf-1-mrj-10.1177_00222437221130722 - Supplemental material for Discovering Online Shopping Preference Structures in Large and Frequently Changing Store Assortments

Supplemental material, sj-pdf-1-mrj-10.1177_00222437221130722 for Discovering Online Shopping Preference Structures in Large and Frequently Changing Store Assortments by Min Kim and Jie Zhang in Journal of Marketing Research

Footnotes

Acknowledgments

This article is based on one of the dissertation essays of the first author. The authors would like to thank an anonymous e-commerce company for providing the data used in this research and Professor Lily Du of Xi’dian University for her help acquiring the data. They are also grateful for the valuable input from Bruno Jacobs, Michel Wedel, and Yogesh Joshi of the University of Maryland, and from seminar participants at the University of Maryland, National University of Singapore, Rutgers University, Vrije Universiteit, University of Technology Sydney, and Korea Advanced Institute of Science and Technology.

Associate Editor

Neeraj Arora

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: The first author gratefully acknowledges financial supports of the American Statistical Association through its Doctoral Research Award in Marketing and the Marketing Science Institute through its Alden Clayton Dissertation Proposal Competition Award (Honorable Mention).

ORCID iD

Min Kim

Notes

References

Adomavicius

Gediminas

Tuzhilin

Alexander

(2005), “Toward the Next Generation of Recommender Systems: A Survey of the State-of-the-Art and Possible Extensions,” IEEE Transactions on Knowledge and Data Engineering, 17 (6), 734–49

Airoldi

Edoardo M.

Blei

David M.

Erosheva

Elena A

Fienberg

Stephen E.

(2014), “Introduction to Mixed Membership Models and Methods,” in Handbook of Mixed Membership Models and Their Applications, Edoardo M. Airoldi, David Blei, Elena A. Erosheva, and Stephen E. Fienberg, eds. CRC Press, 3–14.

Ansari

Asim

Yang

Zhang

Jonathan Z.

(2018), “Probabilistic Topic Model for Hybrid Recommender Systems: A Stochastic Variational Bayesian Approach,” Marketing Science, 37 (6), 987–1008.

Atchison

Shen

Sheng M.

(1980), “Logistic-Normal Distributions: Some Properties and Uses,” Biometrika, 67 (2), 261–72.

Blei

David M.

Lafferty

John D.

(2007), “A Correlated Topic Model of Science,” Annals of Applied Statistics, 1 (1), 17–35.

Blei

David M.

Andrew Y.

Jordan

Michael I.

(2003), “Latent Dirichlet Allocation,” Journal of Machine Learning Research, 3, 993–1022.

Boughanmi

Khaled

Ansari

Asim

(2021), “Dynamics of Musical Success: A Machine Learning Approach for Multimedia Data Fusion,” Journal of Marketing Research, 58 (6), 1034–57.

Bucklin

Randolph E.

Sismeiro

Catarina

(2003), “A Model of Web Site Browsing Behavior Estimated on Clickstream Data,” Journal of Marketing Research, 40 (3), 249–67.

Dew

Ryan

Ansari

Asim

Yang

(2020), “Modeling Dynamic Heterogeneity Using Gaussian Processes,” Journal of Marketing Research, 57 (1), 55–77.

10.

Dotson

Marc R.

Büschken

Joachim

Allenby

Greg M.

(2020), “Explaining Preference Heterogeneity with Mixed Membership Modeling,” Marketing Science, 39 (2), 461–76.

11.

Fader

Peter S.

Hardie

Bruce G.S.

(1996), “Modeling Consumer Choice Among SKUs,” Journal of Marketing Research, 33 (4), 442–52.

12.

Gelman

Andrew

Carlin

John B.

Stern

Hal S.

Dunson

David B.

Vehtari

Aki

Rubin

Donald B.

(2013), Bayesian Data Analysis, 3rd ed. CRC Press.

13.

Gelman

Andrew

Hwang

Jessica

Vehtari

Aki

(2014), “Understanding Predictive Information Criteria for Bayesian Models,” Statistics and Computing, 24 (6), 997–1016.

14.

Gilbride

Timothy

Allenby

Greg M.

(2004), “A Choice Model with Conjunctive, Disjunctive, and Compensatory Screening Rules,” Marketing Science, 23 (3), 391– 406.

15.

Groth

John C.

McDaniel

Stephen W.

(1993), “The Exclusive Value Principle: The Basis for Prestige Racing,” Journal of Consumer Marketing, 10 (1), 10–16.

16.

Guadagni

Peter M.

Little

John D.C.

(1983), “A Logit Model of Brand Choice Calibrated on Scanner Data,” Marketing Science, 2 (3), 203–38.

17.

Hauser

John R.

Wernerfelt

Birger

(1990), “An Evaluation Cost Model of Consideration Sets,” Journal of Consumer Research, 16 (4), 393–408.

18.

Teck-Hua

Chong

Juin-Kuan

(2003), “A Parsimonious Model of Stockkeeping-Unit-Choice,” Journal of Marketing Research, 40 (3), 351–65.

19.

Hoffman

Matthew D.

Gelman

Andrew

(2014), “The No-U-Turn Sampler: Adaptively Setting Path Lengths in Hamiltonian Monte Carlo,” Journal of Machine Learning Research, 15 (1), 1593–1623.

20.

Inman

J. Jeffrey

Park

Joonwook

Sinha

Ashish

(2008), “A Dynamic Choice Map Approach to Modeling Attribute-Level Varied Behavior Among Stock-Keeping Units,” Journal of Marketing Research, 45 (1), 94–103.

21.

Jacobs

Bruno J.D.

Donkers

Bas

Fok

Dennis

(2016), “Model-Based Purchase Predictions for Large Assortments,” Marketing Science, 35 (3), 389–404.

22.

Jacobs

Bruno J.D.

Fok

Dennis

Donkers

Bas

(2021), “Understanding Large-Scale Dynamic Purchase Behavior,” Marketing Science, 40 (5), 844–70.

23.

Kim

Dongwoo

Alice

(2017), “Hierarchical Dirichlet Scaling Process,” Machine Learning, 106, 387–418.

24.

Kök

Gürhan

Fisher

Marshall L.

Vaidyanathan

Ramnath

(2008), “Assortment Planning: Review of Literature and Industry Practice,” Retail Supply Chain Management, 122, 99–152.

25.

Kumar

Uma

Kumar

Vinod

Kapur

J.N.

(1986), “Normalized Measures of Entropy,” International Journal of General Systems, 12 (1), 55–69.

26.

Lee

Hsiao-Ching

Chen

Wei-Wei

Wang

Chih-Wei

(2015), “The Role of Visual Art in Enhancing Perceived Prestige of Luxury Brands,” Marketing Letters, 26 (4), 593–606.

27.

Liu

Jia

Toubia

Oliver

(2018), “A Semantic Approach for Estimating Consumer Content Preferences from Online Search Queries,” Marketing Science, 37 (6), 855–1052.

28.

Moe

Wendy W

. (2003), “Buying, Searching, or Browsing: Differentiating Between Online Shoppers Using In-Store Navigational Clickstream,” Journal of Consumer Psychology, 13 (1–2), 29–39.

29.

Moe

Wendy W

. (2006), “An Empirical Two-Stage Choice Model with Varying Decision Rules Applied to Internet Clickstream Data,” Journal of Marketing Research, 43 (4), 680–92.

30.

Montgomery

Alan L.

Shibo

Srinivasan

Kannan

Liechty

John C.

(2004), “Modeling Online Browsing and Path Analysis Using Clickstream Data,” Marketing Science, 23 (4), 579–95.

31.

Okui

Tasuku

(2020), “A Bayesian Nonparametric Topic Model for Microbiome Data Using Subject Attributes,” IPSJ Transactions on Bioinformatics, 13, 1–6.

32.

Paisley

John

Wang

Chong

Blei

David

(2011), “The Discrete Infinite Logistic Normal Distribution for Mixed-Membership Modeling,” in Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, 74–82.

33.

Roberts

John H.

Lattin

James M.

(1991), “Development and Testing of a Model of Consideration Set Composition,” Journal of Marketing Research, 28 (4), 429–41.

34.

Rooderkerk

Robert P.

van Heerde

Harald J.

Bijmolt

Tammo H.A.

(2013), “Optimizing Retail Assortments,” Marketing Science, 32 (5), 699–715.

35.

Schweidel

David A.

Moe

Wendy W.

(2016), “Binge Watching and Advertising,” Journal of Marketing, 80 (5), 1–19.

36.

Sharif-Razavian

Narges

Zollmann

Andreas

(2008), “An Overview of Nonparametric Bayesian Models and Applications to Natural Language Processing,” Science, 71–93.

37.

Shi

Savannah Wei

Zhang

Jie

(2014), “Usage Experience with Decision Aids and Evolution of Online Purchase Behavior,” Marketing Science, 33 (6), 871–82.

38.

Singh

Vishal P.

Hansen

Karsten T.

Gupta

Sachin

(2005), “Modeling Preferences for Common Attributes in Multicategory Brand Choice,” Journal of Marketing Research, 42 (2), 195–209.

39.

Sinha

Ashish

Inman

Jeffrey

Wang

Yantao

Park

Joonwook

(2005), “Attribute Drivers: A Factor Analytic Choice Map Approach for Understanding Choices Among SKUs,” Marketing Science, 24 (3), 351–66.

40.

Sinha

Ashish

Sahgal

Anna

Mathur

Sharat K.

(2013), “Category Optimizer: A Dynamic-Assortment, New-Product-Introduction, Mix-Optimization, and Demand-Planning System,” Marketing Science, 32 (2), 221–28.

41.

Smith

Michael D.

Brynjolfsson

Erik

(2001), “Consumer Decision-Making at an Internet Shopbot: Brand Still Matters,” Journal of Industrial Economics, 49 (4), 541–58.

42.

Sproles

George B

. (1981), “Analyzing Fashion Life Cycles: Principles and Perspectives,” Journal of Marketing, 45 (4), 116–24.

43.

Stan Development Team (2018), “Stan Modeling Language Users Guide and Reference Manual,” version 2.18, http://mc-stan.org.

44.

Teh

Yee Whye

Jordan

Michael I.

Beal

Matthew J.

Blei

David M.

(2006), “Hierarchical Dirichlet Processes,” Journal of the American Statistical Association, 101 (476), 1566–81.

45.

Tibshirani

Robert

(1996), “Regression Shrinkage and Selection via the Lasso,” Journal of the Royal Statistical Society Series B (Methodological), 58 (1), 267–88.

46.

Tirunillai

Seshadri

Tellis

Gerard J.

(2014), “Mining Meaning from Online Chatter: Strategic Brand Analysis of Big Data Using Latent Dirichlet Allocation,” Journal of Marketing Research, 51 (4), 463–79.

47.

Toubia

Oliver

Iyengar

Garud

Bunnell

Renèe

Lemaire

Alain

(2019), “Extracting Features of Entertainment Products: A Guided Latent Dirichlet Allocation Approach Informed by the Psychology of Media Consumption,” Journal of Marketing Research, 56 (1), 18–36.

48.

Watanabe

Sumio

(2010), “Asymptotic Equivalence of Bayes Cross Validation and Widely Applicable Information Criterion in Singular Learning Theory,” Journal of Machine Learning Research, 11, 3571–94.

49.

Yunjie

Kim

Hee-Woong

(2008), “Order Effect and Vendor Inspection in Online Comparison Shopping,” Journal of Retailing, 84 (4), 477–86.

50.

Zolaktaf

Zainab

Babanezhad

Reza

Pottinger

Rachel

(2018), “A Generic Top-N Recommendation Framework for Trading-Off Accuracy, Novelty, and Coverage,” in Proceedings of the 2018 IEEE 34th International Conference on Data Engineering, 149–60.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.94 MB